Hippocampus supports multi-task reinforcement learning under partial observability.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Biology

Background:

• The hippocampus is linked to navigation, but its precise function in real-world learning is not fully understood.
• Understanding hippocampal function is key to explaining goal-oriented navigation and survival strategies.

Purpose of the Study:

• To investigate the role of hippocampal circuitry in reinforcement learning (RL) for navigation in partially observable environments.
• To model animal behavior and neural data using deep RL agents.

Main Methods:

• Trained deep RL agents in partially observable environments using egocentric and allocentric tasks.
• Compared agents with recurrent hippocampal circuitry to feedforward networks.
• Used dimensionality reduction on agent representations and validated against rat hippocampal recordings.

Main Results:

• Agents with hippocampal circuitry learned tasks consistent with animal behavior, unlike feedforward networks.
• Extracted representations (reward, strategy, temporal) from agents mirrored experimental findings.
• Hippocampal RL agents showed better generalization to new conditions and predicted state-specific trajectories.

Conclusions:

• Recurrent hippocampal networks are crucial for effective RL in naturalistic, partially observable environments.
• The findings support a significant role for the hippocampus in facilitating adaptive navigation and learning.
• This research bridges computational modeling with empirical neuroscience to explain hippocampal function.